HOUSE_OVERSIGHT_018423.jpg

even more symphonies they’d probably be great too. Unfortunately he’s dead. Wouldn't it be nice if we could sample his old symphonies and make new ones whenever we want?2°° In the future we’ll invite Al into our lives to harmonize away many of the problems we face, not merely making up for Mozart’s inconvenient mortality. “AI Agents” will linger along side us. They will compose versions of themselves we'll not quite grasp, even as we appreciate their efficient magic. “Al is both freedom from programming and freedom from understanding,” runs one programmer’s line2®!. Today machines that once demanded millions of lines of code can function with a fraction of that. Instructions are sent to machine learning systems; the programs do the rest. Such designs balance their mystery with efficacy. They speak to and learn from each other too. Part of the reason that the the “Does it think like a human?” Turing Test will be insufficient in the future is that the machines are not learning from only from humans. They are learning from each other. Perhaps this is not such a bad thing. The distinguished physicists George Ellis and Joe Silk, who spent a lifetime trying to stand on Newton and Einstein’s shoulders to grasp answers about gravity or the future of our universe, electrified many of their peers in 2015 with by wondering if perhaps too much of science had become unscientific, unverifiable, unreliable. The great grand ideas of our day, notions like string theory or dark matter, differ in a crucial way from Newton’s laws of motion or Einstein’s principles: They cannot seem to be tested and significantly proved. And this had fired a trend among younger physicits: Perhaps there was no need for proof. To Ellis and Silk this seemed an awful retreat, dragging physics back to a pre- Enlightenment age of conjecture, superstition and instinct. “This year, debates in physics circles took a worrying turn,” they wrote. “Faced with difficulties in applying fundamental theories to the observed Universe, some researchers called for a change in how theoretical physics is done. They began to argue — explicitly — that if a theory is sufficiently elegant and explanatory, it need not be tested.” Fans of such an approach called the idea “post-empirical science.” This strange, oxymoronic idea was, in a sense, like proposing post-rules baseball: A recipie for wild, swinging chaos that would make scorekeeping impossible. The strange, boiling debate did however reflect an underlying and unnerving truth: Science does seem to have stalled. And it became inevitable to ask: Might it be possible that the machines - or some fusion of Shalosh B. Ekhad andta human mind — can reach into an understanding of laws that no human alone can fathom. We’ve said again and again: Connection changes the nature of an object. Perhaps it changes 260 Wouldn't it be nice: Andrej Karpathy, “The Unreasonable Effectiveness of Recurrent Neural Networks,” in The Hackers Guide to Neural Networks published online May 21, 2015 or John Supko, “How I Taught My Computer to Write Its Own Music,” in Nautilus, February 12, 2015 and Daniel Johnson, “Composing Music with Recurrent Neural Networks,” on Hexahedria Blog August 3, 2015 261 Freedom from understanding: Philip Greenspun, “Big data and machine learning” from Philip Greenspun weblog (November 21, 2015) 191 HOUSE_OVERSIGHT_018423